DeepSeek released the Prover-V2 model, with a parameter count reaching 671 billion.

Today, DeepSeek released a new model named DeepSeek-Prover-V2-671B on the AI open-source Community Hugging Face. It is reported that DeepSeek-Prover-V2-671B uses a more efficient safetensors file format and supports multiple computing precisions, making model training and deployment faster and more resource-efficient, with parameters reaching 671 billion, or an upgraded version of last year's Prover-V1.5 mathematical model. In terms of model architecture, this model utilizes the DeepSeek-V3 architecture, adopting the MoE (Mixture of Experts) mode, featuring 61 layers of Transformer layers and a hidden layer of 7168 dimensions. It also supports ultra-long context, with a maximum positional embedding of 163840, allowing it to handle complex mathematical proofs, and it employs FP8 Algo to reduce model size and improve inference efficiency. (Sina Technology)